Nicole Königstein — author of O’Reilly’s forthcoming books on transformers and AI agents, and founder of AgensFlow — answers a question the industry mostly answers too loosely: what would it actually take to call an agent autonomous?

Her definition

  • Everything called an autonomous agent today makes decisions inside a framework that someone else built and fixed in place.
  • A genuinely autonomous system, for her, can redesign itself — see its own errors and improve from them.
  • That means the agent can change the architecture itself, including how the agents inside it communicate with each other, not just its behaviour inside existing rails.
  • Her example of the bar in practice: a system that notices one of its agents contributes nothing, and cuts that agent out of the system.

Why the distinction is worth keeping

  • It separates “makes decisions within bounds” from “owns its own structure” — and the second is what self-improvement actually requires.
  • It is a useful test for agent product claims: ask who is allowed to change the architecture, and whether anything in the loop is even permitted to try.
  • Most autonomy discourse collapses the two, which is how a workflow with a router ends up described as a self-improving system.

“Fully autonomous in my mind would mean we give the agent a way of redesigning themselves.”